developing visual analytics applications: lessons learned ... › gvu › ii › talks ›...

Developing Visual Analytics Applications: Lessons Learned from the Trenches

John Stasko & Carsten Görg

Information Interfaces Research Group

School of Interactive Computing

Georgia Institute of Technology

Dec. 3, 2009

Jigsaw

• Visualization for Investigative Analysis and Sense-making across Document Collections

“Putting the pieces together”

2

3

Blogs

DBs

Documents/case reports

Example Documents

4

Our Focus

• Entities within the documents

– Person, place, organization, phone number, date, license plate, etc.

• Thesis: A bigger story or insight from the documents will involve a set of interconnected entities working in coordination

• (We use entity identification software from others)

5

Connections

• Entities relate/connect to each other to make a larger “story”

• Connection definition:

– Two entities are connected if they appear in a document together

– The more documents they appear in together, the stronger the connection

6

Jigsaw

• Multiple visualizations (views) of documents, entities, & their connections

• Views are highly interactive and coordinated

• User actions generate eventsthat are transmitted to and (possibly) reflected in other views

7

System Views

8

The Need for Pixels

9

Even More!

10

Even More!

11

Demo Data

• NSF awards from 2000-now (abstract & key meta-data)

• Big thanks to Remco Chang & UNCC folks

• Demo focus: CSE awards (12,243)

12

13

14

15

Demo

What We’re Working On

• Jigsaw is heavy on the user-directed visual exploration, but light on the automated computational analysis

16

Recommend Related Entities

17

Adding Computational Analysis

• Document themes & clustering

18

19

Adding Computational Analysis

• Document themes & clustering

• Document similarity

20

21

Similarity

Most similar

Adding Computational Analysis

• Document themes & clustering

• Document similarity

• Document summarization

22

23

Earthquake hits Pacific island of Vanuatu.

Adding Computational Analysis

• Document themes & clustering

• Document similarity

• Document summarization

• Sentiment analysis

24

25

Most positive

Most negative

Lessons Learned

• A quick recap of some things we’ve learned from working with partners from

– Homeland security

– Law enforcement

– Intelligence analysis

– Investigative reporting

– Business intelligence

and from our experience developing systems like Jigsaw

26

Focus on the user’s task & goals

27

But in many instances, you don’t know the questions

28

As the data scales up, we still want interactive visualization

29

Online algorithms are better than offline algorithms

– Real-time is great

30

For real systems, pragmatics matter

– “Common” programming languages

– Others must be able to access your code

– Documentation, tutorials, …

– “Good” is usually good enough, especially at first (great would be wonderful)

31

You don’t get training data

32

Real data doesn’t necessarily have nice clusters

33

Evaluation doesn’t necessarily have a good quantitative measure

34

35

Fundamentalalgorithmicresearch

Application-drivenproblems

Acknowledgments

• Work conducted as part of the Southeastern Regional Visualization and Analytics Center, supported by DHS and NVAC and the new DHS Center of Excellence in Command, Control & Interoperability (VACCINE Center)

• Supported by NSF IIS-0414667, CCF-0808863 (FODAVA lead), NSF IIS-0915788

36